nvidia/Gemma-4-31B-IT-NVFP4
| benchmark | category | score | context | conf | evidence |
|---|---|---|---|---|---|
| AIME 2025 | math | 87.92 | AIME 2025 | medium | | AIME 2025 | 87.92% | 87.60% | |
| GPQA Diamond | reasoning | 85.8 | GPQA Diamond | medium | | GPQA Diamond | 85.80% | 85.35% | |
| MMLU-Pro | knowledge | 85.25 | MMLU Pro | medium | | MMLU Pro | 85.25% | 84.94% | |
| LiveCodeBench | coding | 82.49 | LiveCodeBench (pass@1) | medium | | LiveCodeBench (pass@1) | 82.49% | 82.27% | |
| SciCode | coding | 33.61 | Scicode subtask acc (pass@1) | medium | | Scicode subtask acc (pass@1) | 33.61% | 33.18% | |
| Terminal-Bench | agentic-coding | 27.08 | Terminal-Bench Hard (pass@1) | medium | | Terminal-Bench Hard (pass@1) | 27.08% | 27.08% | |
| title | author | status | comments | type |
|---|---|---|---|---|
| Update chat_template.jinja | xdavxd | closed | 0 | disc |
| fix(chat_template): backport latest Google canonical template | taniguchimasaya | open | 0 | disc |
| Model tree shows "Finetuned" instead of "Quantized" | gaku-gaku | open | 0 | disc |
| How to run this model YAML with sparkrun on NVIDIA GB10 | Sunnypang | open | 0 | disc |
| Update tokenizer_config.json | zhiyucheng | merged | 0 | disc |
| Update chat_template.jinja | zhiyucheng | merged | 0 | disc |
| Update README.md | tstarkey-nvidia | merged | 0 | disc |
| Request: DOI | brayden-peterson | closed | 0 | disc |
| This model wasn't trained with FP4 or NVFP4 | yangus87 | open | 0 | disc |
| 1*H100 with vLLM 0.19.0 Failed | JeffreySheng | open | 0 | disc |
| Question about q_scale / KV cache scale fallback in vLLM for Gemma-4-31B-IT-NVFP4: expected accuracy impact? | Shaoqing | open | 0 | disc |
| Why not quantize the MATRICES of Wq, Wk, Wv, Wo? | BeetSoup | open | 0 | disc |
| 这个版本对于5090单卡来说还是太大了 | iwaitu | open | 0 | disc |
| Why is this 4bit version has a 32.7 GB size? | alexcardo | open | 0 | disc |
| Inference on single RTX5090? | SyferO | closed | 0 | disc |
| Update README.md | tstarkey-nvidia | closed | 0 | disc |